Lately, in the world of artificial intelligence, it seems like we're living in a horse race with only three clear participants. The debate almost always revolves around whether the latest OpenAI's GPT model is better than Claude 3 Opus, or if Google's Gemini has finally managed to catch up. We obsess over benchmarks, context windows, and, of course, monthly subscriptions.

But while we stare intently at Silicon Valley, something fascinating is happening in the East. It's not a simple copy; it's a real alternative that is forcing many early adopters and, above all, technology directors, to turn their heads.

I'm talking about DeepSeek, a set of AI models developed by DeepSeek AI, a startup based in Beijing that, by the way, has ties to Tsinghua University (the "MIT of China"). And the reason you should know about it isn't just its origin, but because it's attacking the only weak point that the AI giants seemed to have well secured: the price.

What exactly is DeepSeek and why all the fuss?

At first glance, DeepSeek is "just another one" in the long list of LLMs (Large Language Models) that have emerged in the last two years. But if you dig a little, you realize it's not.

Its latest flagship model, DeepSeek-V2, launched in 2024, isn't a model for playing at making poems (which it can also do), but an optimized beast. It's a MoE (Mixture of Experts) model, an architecture similar to what Mistral uses in its acclaimed Mixtral 8x7B. Without getting into an engineering class, this means the model is much more efficient: instead of activating its entire massive neural network to respond "hello," it only activates the necessary "parts" (experts).

The result? Performance that, according to its own benchmarks (and those of third parties starting to validate them), rubs shoulders with and, in some tasks, surpasses GPT-4 Turbo and Claude 3 Sonnet.

Where it truly shines, and this is key, is in coding and mathematical reasoning tasks. Its specialized model, DeepSeek-Coder, has become a kind of secret weapon for many developers, often outperforming GitHub Copilot (based on OpenAI) in generating complex code and debugging.

But, let's be honest, there are starting to be many good models. The real mental "click" happens when you look at the next line on its product page.

The elephant in the room: The price is ridiculously low

This is where the story gets interesting. We're used to high-quality AI being expensive. If you've tried to set up a RAG (Retrieval-Augmented Generation) system for your company, or simply wanted to process thousands of documents through OpenAI's API, you know that the API bill at the end of the month can give you a scare.

While many continue to bet on GPT-4 for their operations, the cost is a real brake on massive experimentation.

DeepSeek has decided it doesn't want to play that game. Its API prices aren't a "discount." They're a demolition.

We're talking about figures that, in some cases, are 98% or 99% cheaper than GPT-4 Turbo's. It's not a typo. While processing a million tokens (the "currency" of AI) in an OpenAI model can cost you several dollars, DeepSeek offers figures around cents. Launch offers that sound like science fiction have been seen, such as "one yuan (about 13 euro cents) per million tokens."

Suddenly, the economic barrier to entry for AI collapses.

The "good, nice, and cheap" AI... Where's the catch?

As good engineers and technicians, we know that when something is too good to be true, it usually has a downside. And DeepSeek is no exception. It has several, and they're important.

1. The geopolitical factor (The big "BUT") We can't ignore it: DeepSeek is a Chinese company. For a startup in Seville that wants to analyze sentiment on social media, this may be irrelevant. But for a medium or large company that handles customer data (GDPR), financial information, or industrial secrets, using an API whose traffic passes through servers in China (or is simply under Chinese jurisdiction) is a resounding "no" for compliance and security reasons. Privacy and data sovereignty are, right now, the biggest brake on its adoption in the West at a serious corporate level.

2. The ecosystem and documentation OpenAI, Google, and Meta (with Llama) don't just offer a model; they offer an ecosystem. They have millions of users, forums overflowing with questions and answers, YouTube tutorials, polished SDKs in all languages, and native integration on platforms like Azure or AWS.

DeepSeek, although it has surprisingly good technical documentation, is still the "new kid." Finding help on Stack Overflow is harder. Integrations are more manual. If you get stuck, you're more alone.

3. Is it really as good as they say? Although the benchmarks in coding and mathematics are stellar, some users report that in more "human" tasks—like creative writing, nuance, irony, or "common sense" reasoning—it can still feel a step behind the finesse of a GPT-4o or a Claude 3 Opus. It's not bad at all, but its optimization towards the technical side is noticeable.

The perfect "lab": Savings for experimenting

So then, what is DeepSeek good for if it has these drawbacks?

It's good for the most important thing this sector needs right now: experimenting without going broke.

Imagine you're an IT department (like us at ForgeNEX) and you want to test an idea. For example, creating an internal chatbot that reads the 20,000 pages of technical documentation of all your projects so new employees can ask it questions.

• The plan with GPT-4: You do a small proof of concept (PoC), it works, but when you calculate the cost of indexing (embedding) all those documents and the cost of daily queries from 50 employees, the project goes to thousands of euros per month. The CFO gives you a bad look and the idea stays in a drawer.
• The plan with DeepSeek: You do the PoC. You calculate the cost. And you realize that the total API cost is so low it's almost a rounding error in the department's budget. You don't care if 50 or 100 employees use it. The cost is manageable.

DeepSeek thus becomes the perfect "lab." It's the ideal tool for prototyping, for testing, for failing, and for iterating. It allows you to build that internal log analysis tool, that RAG system for your legal documents, or that script that automatically summarizes GitHub commits, without having to ask accounting for permission every time you run the program.

A new standard or a regional giant?

DeepSeek is proof that excellence in AI isn't a monopoly of Silicon Valley and, above all, that the current price of AI isn't set in stone.

It's possible that, due to the geopolitical factor, DeepSeek may never become the engine of AI in European or American companies. But its mere existence is already great news for everyone. It's putting pressure on OpenAI, Google, and Anthropic, forcing them to optimize their own models and lower their prices to not lose the growing mass of developers and startups simply looking for the best power/price ratio.

You might not use DeepSeek for your final customer-facing product, but it's undoubtedly the tool you want in your toolbox for tinkering. Here at ForgeNEX, we're keeping a close eye on it. In the AI era, the ability to experiment fast and cheap is everything.